The demonstration begins with a simple hosted agent, then builds a research desk that sends user events to a session and forwards its event stream to a browser interface. An environment defines sandbox dependencies and network access, while a skill describes how to retrieve and interpret public SEC filings.
The application adds specialized subagents, a shared memory store and structured research deliverables. Outcome definitions combine a task, evaluation rubric and iteration limit so a separate reviewer can check the result; the workshop presents this as a supervision mechanism, not proof of error-free financial research.
The presenters explain domain-scoped credential substitution, MCP integration, self-hosted sandbox options and portable memory access. They also acknowledge limitations: session-level outcomes do not yet define separate subagent outcomes, budget controls are described as exploratory, and rewind or session-forking behavior is not demonstrated as an available feature. The financial examples illustrate software architecture rather than investment advice.
Watch on YouTube
